Evaluating Sparse Autoencoders on Targeted Concept Erasure Tasks

• URL: http://arxiv.org/abs/2411.18895v1
• Date: Thu, 28 Nov 2024 03:58:48 GMT
• Title: Evaluating Sparse Autoencoders on Targeted Concept Erasure Tasks
• Authors: Adam Karvonen, Can Rager, Samuel Marks, Neel Nanda,
• Abstract summary: Sparse Autoencoders (SAEs) are an interpretability technique aimed at decomposing neural network activations into interpretable units. We introduce a family of evaluations based on SHIFT, a downstream task from Marks et al. We adapt SHIFT into an automated metric of SAE quality; this involves replacing the human annotator with an LLM. We also introduce the Targeted Probe Perturbation (TPP) metric that quantifies an SAE's ability to disentangle similar concepts.
• Score: 1.4565166775409717
• License:
• Abstract: Sparse Autoencoders (SAEs) are an interpretability technique aimed at decomposing neural network activations into interpretable units. However, a major bottleneck for SAE development has been the lack of high-quality performance metrics, with prior work largely relying on unsupervised proxies. In this work, we introduce a family of evaluations based on SHIFT, a downstream task from Marks et al. (Sparse Feature Circuits, 2024) in which spurious cues are removed from a classifier by ablating SAE features judged to be task-irrelevant by a human annotator. We adapt SHIFT into an automated metric of SAE quality; this involves replacing the human annotator with an LLM. Additionally, we introduce the Targeted Probe Perturbation (TPP) metric that quantifies an SAE's ability to disentangle similar concepts, effectively scaling SHIFT to a wider range of datasets. We apply both SHIFT and TPP to multiple open-source models, demonstrating that these metrics effectively differentiate between various SAE training hyperparameters and architectures.

Related papers

• Sparse Autoencoder Features for Classifications and Transferability [11.2185030332009]
  We analyze Sparse Autoencoders (SAEs) for interpretable feature extraction from Large Language Models (LLMs) Our framework evaluates (1) model-layer selection and scaling properties, (2) SAE architectural configurations, including width and pooling strategies, and (3) the effect of binarizing continuous SAE activations.
  arXiv  Detail & Related papers  (2025-02-17T02:30:45Z)
• Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoders [115.34050914216665]
  Sparse Autoencoders (SAEs) have emerged as a powerful unsupervised method for extracting sparse representations from language models. We introduce a suite of 256 SAEs, trained on each layer and sublayer of the Llama-3.1-8B-Base model, with 32K and 128K features. We assess the generalizability of SAEs trained on base models to longer contexts and fine-tuned models.
  arXiv  Detail & Related papers  (2024-10-27T17:33:49Z)
• Automatically Interpreting Millions of Features in Large Language Models [1.8035046415192353]
  sparse autoencoders (SAEs) can be used to transform activations into a higher-dimensional latent space. We build an open-source pipeline to generate and evaluate natural language explanations for SAE features. Our large-scale analysis confirms that SAE latents are indeed much more interpretable than neurons.
  arXiv  Detail & Related papers  (2024-10-17T17:56:01Z)
• Efficient Dictionary Learning with Switch Sparse Autoencoders [8.577217344304072]
  We introduce Switch Sparse Autoencoders, a novel SAE architecture aimed at reducing the compute cost of training SAEs. Inspired by sparse mixture of experts models, Switch SAEs route activation vectors between smaller "expert" SAEs. We find that Switch SAEs deliver a substantial improvement in the reconstruction vs. sparsity frontier for a given fixed training compute budget.
  arXiv  Detail & Related papers  (2024-10-10T17:59:11Z)
• SAGE: Scalable Ground Truth Evaluations for Large Sparse Autoencoders [7.065809768803578]
  We introduce SAGE: Scalable Autoencoder Ground-truth Evaluation, a ground truth evaluation framework for SAEs. We demonstrate that our method can automatically identify task-specific activations and compute ground truth features at these points. Our framework paves the way for generalizable, large-scale evaluations of SAEs in interpretability research.
  arXiv  Detail & Related papers  (2024-10-09T21:42:39Z)
• Adapting Segment Anything Model for Unseen Object Instance Segmentation [70.60171342436092]
  Unseen Object Instance (UOIS) is crucial for autonomous robots operating in unstructured environments. We propose UOIS-SAM, a data-efficient solution for the UOIS task. UOIS-SAM integrates two key components: (i) a Heatmap-based Prompt Generator (HPG) to generate class-agnostic point prompts with precise foreground prediction, and (ii) a Hierarchical Discrimination Network (HDNet) that adapts SAM's mask decoder.
  arXiv  Detail & Related papers  (2024-09-23T19:05:50Z)
• Semi-Supervised One-Shot Imitation Learning [83.94646047695412]
  One-shot Imitation Learning aims to imbue AI agents with the ability to learn a new task from a single demonstration. We introduce the semi-supervised OSIL problem setting, where the learning agent is presented with a large dataset of trajectories. We develop an algorithm specifically applicable to this semi-supervised OSIL setting.
  arXiv  Detail & Related papers  (2024-08-09T18:11:26Z)
• SAFE: a SAR Feature Extractor based on self-supervised learning and masked Siamese ViTs [5.961207817077044]
  We propose a novel self-supervised learning framework based on masked Siamese Vision Transformers to create a General SAR Feature Extractor coined SAFE. Our method leverages contrastive learning principles to train a model on unlabeled SAR data, extracting robust and generalizable features. We introduce tailored data augmentation techniques specific to SAR imagery, such as sub-aperture decomposition and despeckling. Our network competes with or surpasses other state-of-the-art methods in few-shot classification and segmentation tasks, even without being trained on the sensors used for the evaluation.
  arXiv  Detail & Related papers  (2024-06-30T23:11:20Z)
• Task-customized Masked AutoEncoder via Mixture of Cluster-conditional Experts [104.9871176044644]
  Masked Autoencoder(MAE) is a prevailing self-supervised learning method that achieves promising results in model pre-training. We propose a novel MAE-based pre-training paradigm, Mixture of Cluster-conditional Experts (MoCE) MoCE trains each expert only with semantically relevant images by using cluster-conditional gates.
  arXiv  Detail & Related papers  (2024-02-08T03:46:32Z)
• Task-Oriented Sensing, Computation, and Communication Integration for Multi-Device Edge AI [108.08079323459822]
  This paper studies a new multi-intelligent edge artificial-latency (AI) system, which jointly exploits the AI model split inference and integrated sensing and communication (ISAC) We measure the inference accuracy by adopting an approximate but tractable metric, namely discriminant gain.
  arXiv  Detail & Related papers  (2022-07-03T06:57:07Z)
• Meta-Generating Deep Attentive Metric for Few-shot Classification [53.07108067253006]
  We present a novel deep metric meta-generation method to generate a specific metric for a new few-shot learning task. In this study, we structure the metric using a three-layer deep attentive network that is flexible enough to produce a discriminative metric for each task. We gain surprisingly obvious performance improvement over state-of-the-art competitors, especially in the challenging cases.
  arXiv  Detail & Related papers  (2020-12-03T02:07:43Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.